03 .. fourier ma .pdf



Nom original: 03-..-fourier-ma.pdf

Ce document au format PDF 1.3 a été généré par 3B2 Total Publishing System 7.51c/W / Acrobat Distiller 5.0.5 (Windows), et a été envoyé sur fichier-pdf.fr le 26/04/2012 à 11:55, depuis l'adresse IP 81.255.x.x. La présente page de téléchargement du fichier a été vue 944 fois.
Taille du document: 2.5 Mo (15 pages).
Confidentialité: fichier public




Télécharger le fichier (PDF)










Aperçu du document


IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 12,

DECEMBER 2003

1519

Personal Identification Based on
Iris Texture Analysis
Li Ma, Tieniu Tan, Senior Member, IEEE, Yunhong Wang, Member, IEEE, and Dexin Zhang
Abstract—With an increasing emphasis on security, automated personal identification based on biometrics has been receiving
extensive attention over the past decade. Iris recognition, as an emerging biometric recognition approach, is becoming a very active
topic in both research and practical applications. In general, a typical iris recognition system includes iris imaging, iris liveness
detection, and recognition. This paper focuses on the last issue and describes a new scheme for iris recognition from an image
sequence. We first assess the quality of each image in the input sequence and select a clear iris image from such a sequence for
subsequent recognition. A bank of spatial filters, whose kernels are suitable for iris recognition, is then used to capture local
characteristics of the iris so as to produce discriminating texture features. Experimental results show that the proposed method has an
encouraging performance. In particular, a comparative study of existing methods for iris recognition is conducted on an iris image
database including 2,255 sequences from 213 subjects. Conclusions based on such a comparison using a nonparametric statistical
method (the bootstrap) provide useful information for further research.
Index Terms—Iris recognition, image quality assessment, multichannel spatial filters, texture analysis, biometrics.

æ
1

INTRODUCTION

T

HE recent advances of information technology and the
increasing requirement for security have led to a rapid
development of intelligent personal identification systems
based on biometrics. Biometrics [1], [2] employs physiological or behavioral characteristics to accurately identify each
subject. Commonly used biometric features include face,
fingerprints, voice, facial thermograms, iris, retina, gait,
palm-prints, hand geometry, etc. [1], [2]. Of all these
biometric features, fingerprint verification has received
considerable attention and has been successfully used in
law enforcement applications. Face recognition and speaker
recognition have also been widely studied over the last
25 years, whereas iris recognition is a newly emergent
approach to personal identification [1], [2]. It is reported in
[3] that iris recognition is one of the most reliable
biometrics.
The human iris, an annular part between the pupil
(generally, appearing black in an image) and the white
sclera as shown in Fig. 8, has an extraordinary structure and
provides many interlacing minute characteristics such as
freckles, coronas, stripes, etc. These visible characteristics,
which are generally called the texture of the iris, are unique
to each subject [5], [6], [7], [12], [15], [16], [17], [18], [19], [20],
[21]. Individual differences that exist in the development of
anatomical structures in the body result in such uniqueness.
Compared with other biometric features (such as face,

voice, etc.), the iris is more stable and reliable for
identification [1], [2], [3]. Furthermore, since the iris is an
externally visible organ, iris-based personal identification
systems can be noninvasive to their users [12], [16], [17],
[18], [19], [20], [21], which is of great importance for
practical applications. All these desirable properties (i.e.,
uniqueness, stability, and noninvasiveness) make iris
recognition a particularly promising solution to security in
the near future.
A typical iris recognition system is schematically shown
in Fig. 1. It involves three main modules.
.

.
. The authors are with the National Laboratory of Pattern Recognition,
Institute of Automation, Chinese Academy of Sciences, PO Box 2728,
Beijing, P.R. China, 100080.
E-mail: {lma, tnt, wangyh, dxzhang}@nlpr.ia.ac.cn.
Manuscript received 16 Oct. 2002; revised 31 May 2003; accepted 6 Aug.
2003.
Recommended for acceptance by M. Pietikainen.
For information on obtaining reprints of this article, please send e-mail to:
tpami@computer.org, and reference IEEECS Log Number 117595.
0162-8828/03/$17.00 ß 2003 IEEE

Image acquisition. It is to capture a sequence of iris
images from the subject using a specifically designed
sensor. Since the iris is fairly small (its diameter is
about 1 cm) and exhibits more abundant texture
features under infrared lighting, capturing iris
images of high quality is one of the major challenges
for practical applications. Fortunately, much work
has been done on iris image acquisition [9], [10], [11],
[12], [20], [21], which has made noninvasive imaging
at a distance possible. When designing an image
acquisition apparatus, one should consider three
main aspects, namely, the lighting system, the
positioning system, and the physical capture system
[20]. More recent work on iris imaging may be found
on an iris recognition Web site [14].
Iris liveness detection. Being easy to be forged and
used illegally is a crucial weakness of traditional
personal identification methods. Similarly, it is also
possible that biometric features are forged and
illegally used. Iris liveness detection aims to ensure
that an input image sequence is from a live subject
instead of an iris photograph, a video playback, a
glass eye, or other artifacts. However, efforts on iris

Published by the IEEE Computer Society

1520

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 12,

DECEMBER 2003

Fig. 1. Block diagram of a typical iris recognition system.

liveness detection are still limited, though iris
liveness detection is highly desirable. Daugman [1],
[19], and Wildes [21] mentioned this topic in their
work, but they only described some possible
schemes and did not document a specific method.
How to utilize the optical and physiological characteristics of the live eye to implement effective
liveness detection remains to be an important
research topic.
. Recognition. This is the most key component of an
iris recognition system and determines the system’s
performance to a large extent. Iris recognition
produces the correct result by extracting features of
the input images and matching these features with
known patterns in the feature database. Such a
process can be divided into four main stages: image
quality assessment and selection, preprocessing,
feature extraction, and matching. The first stage
solves the problem of how to choose a clear and
well-focused iris image from an image sequence for
recognition. Preprocessing provides an effective iris
region in a selected image for subsequent feature
extraction and matching.
In addition to recognition, our work on iris-based
personal identification also involves iris sensor design [13]
and implementation of a specific method for iris liveness
detection based on the two schemes described by Daugman
[1], [19]. In this paper, we detail a texture analysis-based
recognition method. Experimental results on an iris image
database including 2,255 image sequences from 213 subjects
have demonstrated that the proposed method is highly
feasible and effective for personal identification. The
novelty of this paper includes the following:
In order to select a suitable image from an image
sequence for accurate recognition, an effective
scheme for image quality assessment is proposed.
2. A bank of spatial filters, whose kernels are suitable
for iris recognition, is defined to capture local details
of the iris so as to produce discriminating texture
features.
3. Using a nonparametric statistical approach, extensive performance comparison of existing schemes for
iris recognition is conducted on a reasonably sized
iris database (To the best of our knowledge, this is
the first comparative study on iris recognition).
The remainder of this paper is organized as follows:
Section 2 briefly summarizes related work. A detailed
description of the proposed method for iris recognition is
given in Section 3. Section 4 reports experiments and
results. Section 5 concludes this paper.
1.

2

RELATED WORK

Using iris patterns as an approach to personal identification
and verification goes back to the late 19th century [8], [21],
but most work on iris recognition [9], [10], [11], [12], [13],
[14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25],
[26], [28], [29], [30], [31] is done in the last decade. Existing
methods for iris recognition mainly concentrate on iris
representation and matching which is also one of the
focuses of this paper.
Unlike fingerprints, it is difficult to classify and localize
semantically meaningful features in an iris image. From the
viewpoint of feature extraction, previous iris recognition
methods can be roughly divided into three major categories:
phase-based methods [16], [17], [18], [19], zero-crossing
representation methods [22], [25], and texture-analysisbased methods [20], [23], [28], [29], [30]. Daugman [17],
[18], [19] used multiscale quadrature wavelets to extract
texture phase structure information of the iris to generate a
2,048-bit iriscode and compared the difference between a
pair of iris representations by computing their Hamming
distance. In [24], Sanchez-Reillo and Sanchez-Avila provided a partial implementation of the algorithm by Daugman. Boles and Boashash [22] calculated a zero-crossing
representation of 1D wavelet transform at various resolution levels of a concentric circle on an iris image to
characterize the texture of the iris. Iris matching was based
on two dissimilarity functions. Sanchez-Avila and SanchezReillo [25] further developed the method of Boles and
Boashash by using different distance measures (such as
Euclidean distance and Hamming distance) for matching.
Wildes et al. [20] represented the iris texture with a
Laplacian pyramid constructed with four different resolution levels and used the normalized correlation to determine whether the input image and the model image are
from the same class. Lim et al. [23] decomposed an iris
image into four levels using 2D Haar wavelet transform and
quantized the fourth-level high frequency information to
form an 87-bit code. A modified competitive learning neural
network (LVQ) was used for classification. Our previous
work [28], [29] adopted a well-known texture analysis
method (multichannel Gabor filtering) to capture both
global and local details in an iris. More recently, Tisse et
al. [26] constructed the analytic image (a combination of the
original image and its Hilbert transform) to demodulate the
iris texture. Emergent frequency images used for feature
extraction are in essence samples of the phase gradient
fields of the analytic image’s dominant components [27].
Similar to the algorithm by Daugman, they sampled binary
emergent frequency images to form a feature vector and
used Hamming distance for matching. It should be noted
that all these algorithms are based on gray images, and

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

Fig. 2. The flowchart of our approach.

color information is not used. The main reason is that the
most important information in recognition (i.e., texture
variations of the iris) is the same in both gray and color
images.
A great deal of progress in iris recognition has been
made through these efforts, therefore, a detailed performance comparison of these algorithms is not trivial. Such
evaluation will be discussed in Section 4. In this paper, we
propose a new iris recognition algorithm based on texture
analysis. Of particular concern and importance in this
method are image quality assessment, feature extraction,
and matching. These very key issues will be addressed in
the following section.

3

IRIS RECOGNITION

In our framework, an iris recognition algorithm includes
four basic modules: image quality assessment and selection,
preprocessing, feature extraction, and matching. Fig. 2
shows how the proposed algorithm works. The solid boxes
are the processed data at different stages and the dashed
boxes denote four different processing steps, respectively.
Detailed descriptions of these steps are introduced as
follows.

3.1 Image Quality Assessment and Selection
When capturing iris images, one usually obtains a sequence
of images rather than a single image. Unfortunately, not all
images in the input sequence are clear and sharp enough for
recognition. As shown in the top row of Fig. 3, the second
image is out of focus, the third one contains many
noticeable interlacing lines (especially in regions close to

1521

the boundary) caused by eye motion, and the last one is an
example of severe occlusions by eyelids and eyelashes.
Therefore, it is necessary to select a suitable image of high
quality from an input sequence before all other operations.
Image quality assessment is an important issue of iris
recognition as the quality of an image strongly affects
recognition accuracy. However, efforts on image quality
assessment are still limited. Daugman [18] measured the
total high frequency power in the 2D Fourier spectrum of an
image to assess the focus of the image. If an image can pass
a minimum focus criterion, it will be used for recognition.
However, Daugman did not provide a detailed description
of his method. Zhang and Salganicoff [31] analyzed the
sharpness of the boundary between the pupil and the iris to
determine whether an image is in focus. Both these methods
aim to measure the focus of an iris image. The former is a
common approach to focus detection able to be used in
various applications, whereas the latter considers the
specific properties of the iris image. Here, we present an
effective scheme to assess image quality by analyzing the
frequency distribution of the iris image. Iris images of low
quality can be roughly categorized into three classes,
namely, out-of-focus images (also called defocused images),
motion blurred images, and images severely occluded by
eyelids and eyelashes. When the subject is far from the
focus plane of the camera, a defocused image like Fig. 3b
will form. If the subject moves during imaging, a motion
blurred image as shown in Fig. 3c will result. When the
subject opens his eye partially, the resulting image as
shown in Fig. 3d contains little useful information. These
images often occur in a captured sequence since the eye is in
the state of continual motion and noninvasive image
acquisition also requires users to adjust their position
(hence, body motion).
Here, the region of interest in an image is the iris, and we
thus focus on only two iris subregions in the horizontal
direction as shown in Fig. 3 for further analysis. That is, we
will utilize information of the iris image as much as
possible. From the viewpoint of frequency analysis, the
spectrum of a defocused iris is greatly dominated by low

Fig. 3. Differences between high quality images and low quality images. (a) A clear image. (b) A defocused image. (c) A motion blurred image. (d) An
occluded image. (e) Fourier spectra of two local iris regions denoted by the white boxes in (a). (f), (g), and (h) are the Fourier spectra corresponding
to (b), (c), and (d), respectively. The quality descriptors of (a), (b), (c), and (d) are ½1:65 106 ; 0:94 , ½1:35 106 ; 0:63 , ½1:62 106 ; 0:55 , and
½2:39 106 ; 0:85 , respectively.

1522

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

frequency components, and an iris image having the eye
partially opened contains significant middle and high
frequency components resulted from the heavy eyelashes.
As far as motion blurred images are concerned, there are
two cases. If the iris sensor works in the interlaced scan
mode (i.e., a frame is divided into two fields which are
captured in an interval of 20 ms or less), the resulting image
as shown in Fig. 3c involves obvious interlacing lines
(hereinafter, called aliasing) in the horizontal direction in
the boundary regions. Such aliasing corresponding to
vertical high frequency components in Fourier spectrum is
more noticeable in regions close to the pupil and the
eyelashes because the pupil and the eyelashes generally
stand in high contrast against their surroundings. If the iris
sensor works in the progressive scan mode (i.e., a complete
frame is generated in one time), smearing along the motion
direction instead of serious aliasing will occur in the image.
Different from the former, the second kind of motion
blurred images lacks middle and high frequency components and has frequency distribution similar to that of
defocused images [32]. In our experiments, motion blurred
images belong to the former since our iris sensor only works
in the interlaced scan mode. In comparison with these
images of low quality, a clear and properly focused iris
image has relatively uniform frequency distribution. This
can be observed in Fig. 3. We thus define the following
quality descriptor:


F2
D ¼ ðF1 þ F2 þ F3 Þ;
F1 þ F3
ZZ
ð1Þ
Fi ¼
jF
ðu;
vÞjdudv
i
¼
1;
2;
3;
pffiffiffiffiffiffiffiffiffi
¼fðu;vÞjf1i < u2 þv2 <¼f2i g

where F ðu; vÞ is the 2D Fourier spectrum of an iris region,
F1 , F2 , and F3 are the power of low, middle, and high
frequency components, respectively, f1i and f2i are the
radial frequency pair and bound the range of the
corresponding frequency components. In our experiments,
three frequency pairs of (0, 6), (6, 22), and (22, 32) are used.
The quality descriptor D consists of two discriminating
frequency features. The first feature is the total spectrum
power of an iris region which can effectively discriminate
clear iris images from severely occluded iris images. The
second feature is the ratio of the middle frequency power
to other frequency power. It should be larger for the clear
image than for the defocused and motion blurred image
since the former has much more middle frequency
information. A complete diagram of the proposed algorithm for quality assessment is plotted in Fig. 4. We first
locate two 64 64 iris regions and compute their quality
descriptors, respectively. Then, the mean of the resulting
two local quality descriptors is regarded as an appropriate
quality measure of an iris image. For a given quality
descriptor, the SVM method is used to distinguish whether
the corresponding iris image is clear. Here, because
defocused images, motion blurred images, and occluded
images do not form a compact cluster in the feature space
defined by the quality descriptor, we use the SVM method
to characterize the distribution boundary of the quality
descriptor between low quality images and clear images

VOL. 25, NO. 12,

DECEMBER 2003

Fig. 4. The flowchart of the proposed method for image quality
assessment.

(see Fig. 9). Note that the scheme for coarse localization of
the pupil is the same as that introduced in Section 3.2.1.
Using the above algorithm, one can accurately assess the
quality of an image. Because the input data is an image
sequence in our experiments, we adopt a simple selection
scheme as follows:
Compute the quality descriptor of each image in the
input sequence.
2. Generate a candidate image set where each image
should successfully pass the above quality assessment (or the quality descriptor of each image is close
to the decision boundary).
3. Select the image whose quality descriptor is the
farthest to the decision boundary in all quality
descriptors of the candidate images.
The combination of the quality assessment algorithm and
the image selection scheme makes sure that an iris image of
high quality is identified from the input sequence. In the
following sections, we will focus on iris representation and
matching based on a single image.
1.

3.2 Image Preprocessing
An iris image, as shown in Fig. 5a, contains not only the
region of interest (iris) but also some “unuseful” parts (e.g.,
eyelid, pupil, etc.). A change in the camera-to-eye distance
may also result in variations in the size of the same iris.
Furthermore, the brightness is not uniformly distributed
because of nonuniform illumination. Therefore, before
feature extraction,the original image needs to be preprocessed to localize iris, normalize iris, and reduce the
influence of the factors mentioned above. Such preprocessing is detailed in the following subsections.
3.2.1 Iris Localization
The iris is an annular part between the pupil (inner
boundary) and the sclera (outer boundary). Both the inner
boundary and the outer boundary of a typical iris can
approximately be taken as circles. However, the two circles
are usually not concentric [17]. We localize the iris using the
following simple but effective method.
1.

Project the image in the vertical and horizontal
direction to approximately estimate the center
coordinates ðXp ; Yp Þ of the pupil. Since the pupil is
generally darker than its surroundings, the coordinates corresponding to the minima of the two

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

1523

system. The iris in the new coordinate system can be
represented in a fixed parameter interval. That is, this
method normalizes irises of different size to the same size.
Similar to this scheme, we counterclockwise unwrap the iris
ring to a rectangular block with a fixed size. Such
unwrapping can be denoted as:
In ðX; Y Þ ¼ Io ðx; yÞ
Y
M
Y
y ¼ yp ð Þ þ ððyi ð Þ yp ð ÞÞ
M
¼ 2 X=N;
x ¼ xp ð Þ þ ððxi ð Þ xp ð ÞÞ

ð2Þ

where In is a M N (64 512 in our experiments) normalized image, ðxp ð Þ; yp ð ÞÞ and ðxi ð Þ; yi ð ÞÞ are the coordinates of the inner and outer boundary points in the direction
in the original image Io . The normalization not only
reduces to a certain extent the iris distortion caused by pupil
movement but also simplifies subsequent processing.

Fig. 5. Image preprocessing. (a) Original image. (b) Localized image.
(c) Normalized image. (d) Estimated background illumination.
(e) Normalized image after enhancement.

projection profiles are considered as the center
coordinates of the pupil.
2. Binarize a 120 120 region centered at the point
ðXp ; Yp Þ by adaptively selecting a reasonable threshold using the gray-level histogram of this region. The
centroid of the resulting binary region is considered
as a more accurate estimate of the pupil coordinates.
In this binary region, we can also roughly compute
the radius of the pupil.
3. Calculate the exact parameters of these two circles
using edge detection (Canny operator in experiments) and Hough transform in a certain region
determined by the center of the pupil.
In the above method, the first two steps provide an
approach to coarse localization of the pupil which can be
used in image quality assessment. In experiments, we
perform the second step twice for a reasonably accurate
estimate. Compared with the localization method by
Wildes et al. [20] where the combination of edge detection
and Hough transform is also adopted, our method
approximates the pupil position before edge detection
and Hough transform. This will reduce the region for edge
detection and the search space of Hough transform and,
thus, result in lower computational demands.

3.2.2 Iris Normalization
Irises from different people may be captured in different
size and, even for irises from the same eye, the size may
change due to illumination variations and other factors.
Such elastic deformation in iris texture will affect the results
of iris matching. For the purpose of achieving more accurate
recognition results, it is necessary to compensate for the iris
deformation. Daugman [17], [18], [19] solved this problem
by projecting the original iris in a Cartesian coordinate
system into a doubly dimensionless pseudopolar coordinate

3.2.3 Image Enhancement
The normalized iris image has low contrast and may have
nonuniform brightness caused by the position of light
sources. All these may affect the subsequent processing in
feature extraction and matching. In order to obtain a more
well-distributed texture image, we first approximate intensity variations across the whole image. The mean of each
16 16 small block constitutes a coarse estimate of the
background illumination. This estimate is further expanded
to the same size as the normalized image by bicubic
interpolation. The estimated background illumination as
shown in Fig. 5d is subtracted from the normalized image to
compensate for a variety of lighting conditions. Then, we
enhance the lighting corrected image by means of histogram equalization in each 32 32 region. Such processing
compensates for the nonuniform illumination, as well as
improves the contrast of the image. Fig. 5e shows the
preprocessing result of an iris image, from which we can see
that finer texture characteristics of the iris become clearer
than those in Fig. 5c.
3.3 Feature Extraction
The iris has a particularly interesting structure and provides
abundant texture information. So, it is desirable to explore
representation methods which can capture local underlying
information in an iris. From the viewpoint of texture
analysis, local spatial patterns in an iris mainly involve
frequency and orientation information. Generally, the iris
details spread along the radial direction in the original
image corresponding to the vertical direction in the
normalized image (see Figs. 7 and 8). As a result, the
differences of orientation information among irises seem to
be not significant. That is, frequency information should
account for the major differences of irises from different
people. We thus propose a scheme to capture such
discriminating frequency information which reflects the
local structure of the iris. In general, the majority of useful
information of the iris is in a frequency band of about three
octaves [18]. Therefore, a bank of filters is constructed to
reliably acquire such information in the spatial domain. As

1524

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 12,

DECEMBER 2003

Fig. 6. The responses of the filters defined in (3). (a) The defined filter with x ¼ y . (b) The defined filter with x > y . (c) Gabor filter. (d), (e), and
(f) are the Fourier spectra of (a), (b), and (c), respectively.

we know, coefficients of the filtered image effectively
indicate the frequency distribution of an image. Two
statistic values are thus extracted from each small region
in the filtered image to represent local texture information
of the iris. A feature vector is an ordered collection of all
features from the local regions. More details of this
algorithm are presented as follows.

3.3.1 Spatial Filters
In the spatial domain, one can extract information of an
image at a certain orientation and scale using some specific
filters, such as Gabor filters [33], [34], [35], [36], [37].
Recently, Gabor filter based methods have been widely
used in computer vision, especially for texture analysis [35],
[36], [37]. Gabor elementary functions are Gaussians
modulated by oriented complex sinusoidal functions. Here,
according to the characteristics of the iris texture, we define
new spatial filters to capture local details of the iris. The
difference between Gabor filter and the defined filter lies in
the modulating sinusoidal function. The former is modulated by an oriented sinusoidal function, whereas the latter
by a circularly symmetric sinusoidal function. Their kernels

Fig. 7. ROIs from three iris samples (after preprocessing).

are given as follows (here, we only consider evensymmetric Gabor filters):
"
!#
1
1 x2 y 2
Mi ðx; y; fÞ; i ¼ 1; 2:
exp
þ
Gðx; y; fÞ ¼
2 x y
2 x2 y2
h
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi i
x2 þ y 2 ;
M1 ðx; y; fÞ ¼ cos 2 f
M2 ðx; y; fÞ ¼ cos½2 f ðx cos þ y sin Þ ;
ð3Þ
where Mi ðx; y; fÞ denotes the modulating function, M1 and
M2 are the modulating function of the defined filter and
Gabor filter, respectively, f is the frequency of the
sinusoidal function, x and y are the space constants of
the Gaussian envelope along the x and y axis, respectively,
and denotes the orientation of Gabor filter. For the defined
filter, when x equals to y (i.e., Gaussian function is
isotropic), one can obtain a bandpass filter with a specific
center frequency. When x and y are different, it not only
considers information from every orientation but also

Fig. 8. Iris samples. Images in the first row are from both eyes of two
Chinese, and the first two in the second row are from Chinese and the
last two from French.

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

shows more interest in information in x or y direction
(determined by x and y ). This is greatly different from a
Gabor filter which can only provide information of an
image at a certain orientation. Fig. 6 clearly shows the
differences between a Gabor filter and the defined spatial
filter. As mentioned earlier, local details of the iris generally
spread along the radial direction, so information density in
the angular direction corresponding to the horizontal
direction in the normalized image is higher than that in
other directions, which is validated by our experimental
results in Section 4.3. Thus, we should pay more attention to
useful information in the angular direction. The defined
filter can well satisfy such requirements of iris recognition.
In other words, the defined kernel is suitable for iris
recognition.
In our experiments, we find that the upper portion of a
normalized iris image (corresponding to regions closer to
the pupil) provides the most useful texture information for
recognition (see Fig. 7). In addition, eyelids and eyelashes
rarely occlude this section. So, we extract features only in
this section (called region of interest, ROI) shown as the
region above the dotted line in Fig. 7. As mentioned above,
useful iris information distributes in a specific frequency
range. We therefore use the defined spatial filters in two
channels to acquire the most discriminating iris features. x
and y used in the first channel are 3 and 1.5, and the second
channel 4.5 and 1.5. In a much shorter version of this method
in [30], we vertically divided the ROI into three subregions
of the same size and estimated the energy of each subregion
within a frequency band. These energy measures were used
as features. In contrast, using multiple filters with different
frequency response for the entire ROI can generate more
discriminating features since different irises have distinct
dominant frequencies.This means that the improved scheme
would be more effective than our earlier one [30].

3.3.2 Feature Vector
According to the above scheme, filtering the ROI (48 512)
with the defined multichannel spatial filters results in
ZZ
Fi ðx; yÞ ¼
Iðx1 ; y1 ÞGi ðx x1 ; y y1 Þdx1 dy1 ;
i ¼ 1; 2;
ð4Þ
where Gi is the ith channel of the spatial filters, Iðx; yÞ
denotes the ROI, and Fi ðx; yÞ is the filtered image. To
characterize local texture information of the iris, we
extract statistical features in each 8 8 small block of
the two filtered images. In our experiments, the total
number of small blocks is 768½ð48 512Þ=ð8 8Þ 2 . For
each small block, two feature values are captured. This
generates 1,536 feature components. The feature values
used in the algorithm are the mean m and the average
absolute deviation of the magnitude of each filtered
block defined as


1X
jFi ðx; yÞj;
n w

¼

1X
jjFi ðx; yÞj mj;
n w

ð5Þ

where w is an 8 8 block in the filtered image, n is the
number of pixels in the block w, and m is the mean of

1525

the block w. These feature values are arranged to form a
1D feature vector
V ¼ ½m1 ; 1 ; m2 ; 2 . . . m768 ; 768 T :

ð6Þ

3.4 Iris Matching
After feature extraction, an iris image is represented as a
feature vector of length 1,536. To improve computational
efficiency and classification accuracy, Fisher linear discriminant is first used to reduce the dimensionality of the
feature vector and then the nearest center classifier is
adopted for classification.
Two popular methods for dimensionality reduction are
principal component analysis and Fisher linear discriminant. Compared with principal component analysis, Fisher linear discriminant not only reduces the dimensionality
of features but also increases class separability by
considering both information of all samples and the
underlying structure of each class. This is also the reason
that Wildes et al. [20] adopted Fisher linear discriminant
rather than general distance measures for iris matching
(though the feature vector includes only four components
in their method). Fisher linear discriminant searches for
projected vectors that best discriminate different classes in
terms of maximizing the ratio of between-class to withinclass scatter. Further details of Fisher linear discriminant
may be found in [38], [39].
The new feature vector f can be denoted as:
f ¼ WTV ;

ð7Þ

where W is the projection matrix and V is the original
feature vector derived in Section 3.3.2. The proposed
algorithm employs the nearest center classifier defined in
(8) for classification in a low-dimensional feature space.
m ¼ arg min dn ðf; fi Þ;
1 i c
X

f j f j
d1 ðf; fi Þ ¼

n ¼ 1; 2; 3:

i

j

d2 ðf; fi Þ ¼

X

f j fij

2

ð8Þ

j

d3 ðf; fi Þ ¼ 1

f T fi
;
k f k k fi k

where f and fi are the feature vector of an unknown sample
and the ith class, respectively, f j and fij are the jth
component of the feature vector of the unknown sample
and that of the ith class, respectively, c is the total number
of classes, k k indicates the Euclidean norm, dn ðf; fi Þ
denotes similarity measure, d1 , d2 , and d3 are L1 distance
measure, L2 distance measure (i.e., Euclidean distance) and
cosine similarity measure, respectively. The feature vector f
is classified into the mth class, the closest mean, using
similarity measure dn ðf; fi Þ.
It is desirable to obtain an iris representation invariant to
translation, scale, and rotation. In our algorithm, translation
invariance and approximate scale invariance are achieved
by normalizing the original image at the preprocessing step.
Most existing schemes achieve approximate rotation invariance either by rotating the feature vector before matching
[17], [18], [19], [22], [24], [26], or by registering the input

1526

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 12,

DECEMBER 2003

Fig. 9. The distributions of the quality descriptor for different types of images.

image with the model before feature extraction [20]. Since
features in our method are projection values by feature
reduction, there is no explicit relation between features and
the original image. We thus obtain approximate rotation
invariance by unwrapping the iris ring at different initial
angles. Considering that the eye rotation is not very large in
practical applications,these initial angle values are -9, -6, -3,
0, 3, 6, and 9 degrees. This means that we define seven
templates which denote the seven rotation angles for each
iris class in the database. When matching the input feature
vector with the templates of an iris class, the minimum of
the seven scores is taken as the final matching distance.

4

EXPERIMENTS

This paper presents a new method for identifying individuals from an iris image sequence. We thus perform a series
of experiments to evaluate its performance. Moreover, we
compare the proposed method with some existing methods
for iris recognition and present detailed discussions on the
overall experimental results.
Evaluating the performance of biometric algorithms is a
difficult issue since it is greatly influenced by all sources of
noise (such as sensor noise and environment noise), the
test database and the evaluation method. Obviously, it is
impossible to model all noise sources and build a test data
set including biometric samples from all subjects in the
world. Thus, using modern statistical methods to measure
the performance of biometric algorithms is a desirable
approach. In this paper, the bootstrap [43], which provides
a powerful approach to estimating the underlying distribution of the observed data using computer-intensive
methods, is adopted to estimate the error rates of a
biometric method. The bootstrap can infer how much
variation in performance measures resulted from a limited
data set can be expected in a larger subject population
using confidence intervals of performance measures. The
bootstrap is in nature a nonparametric empirical method.
Given that we have an original sample x including n
observed data fx1 ; x2 ; . . . ; xn g from an unknown probability distribution F , we can empirically estimate the
distribution F and some characteristics of interest ðF Þ
associated with F by the bootstrap. A key step in the

bootstrap is to generate thousands of random samples
x ¼ fx 1 ; x 2 ; . . . ; x n g (called bootstrap samples) with the
same size as the original sample x by drawing with
replacement. Using the resulting bootstrap samples, one
can easily estimate the statistics of interest. More details of
the bootstrap may be found in [40], [41], [43].
We exploit both interval estimation (a confidence
interval) and commonly used point estimation (only a
numerical value) of statistical measures to characterize the
performance of the methods for iris recognition. This means
that the evaluation is more accurate and effective. The
proposed algorithm is tested in two modes: identification
(i.e., one-to-many matching) and verification (i.e., one-toone matching). In identification mode, the algorithm is
measured by Correct Recognition Rate (CRR), the ratio of
the number of samples being correctly classified to the total
number of test samples. In verification mode, the Receiver
Operating Characteristic (ROC) curve is used to report the
performance of the proposed method. The ROC curve is a
False Match Rate (FMR) versus False NonMatch Rate
(FNMR) curve [3], [4] which measures the accuracy of
matching process and shows the overall performance of an
algorithm. The FMR is the probability of accepting an
imposter as an authorized subject and the FNMR is the
probability of an authorized subject being incorrectly
rejected. Points on this curve denote all possible system
operating states in different trade offs.

4.1 Image Database
Unlike fingerprints and face, there is no common iris
database of a reasonable size. Most existing methods for iris
recognition used small image sets for performance evaluation, and only the method by Daugman has been tested on a
larger image set involving over 200 subjects [3], [19].
Currently, there is also no detailed comparison among the
methods in [16], [17], [18], [19], [20], [21], [22], [23], [24], [25],
[26], [28], [29], [30]. So, we construct an iris image database
named CASIA Iris Database to compare their performance
and provide detailed discussions as well. The CASIA Iris
Database includes 2,255 iris image sequences from 213 subjects (note that this is currently the largest iris database we

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

1527

TABLE 1
CASIA Iris Database

can find in the public domain). Each sequence of the CASIA
Iris Database contains 10 frames acquired in about half a
second. All images are captured using a homemade digital
optical sensor [13]. This sensor works in PAL mode (i.e.,
25 frames/second) and provides near infrared illumination
under which the iris exhibits more abundant texture
features. The subject needs to position himself about 4 cm
in front of the sensor to obtain a clear iris image. Moreover,
a surface-coated semitransparent mirror is placed in front of
the lens so that a person can see and keep his eye in the
center of the sensor. The captured iris images are 8-bit gray
images with a resolution of 320 280. In general, the
diameter of the iris in an image from our database is greater
than 200 pixels. This makes sure that there is enough texture
information for reliable recognition. During image acquisition, short-sighted subjects are requested to take off their
eyeglasses to obtain high quality iris images. However,
contact eyewears are an exception. In our database, about
5.2 percent of the subjects wear contacts. The profile of the
database is shown in Table 1. The subjects consist of
203 members of the CAS Institute of Automation and
10 visiting students from Europe.
The CASIA Iris Database is gradually expanded to contain
more images from more subjects. Currently, it is composed
of two main parts. The first one (namely our earlier
database [29]) contains 500 sequences from 25 different
people. Each individual provides 20 sequences (10 for each
eye) captured in two different stages. In the first stage, five
sequences of each eye are acquired. Four weeks later, five
more sequences of each eye are obtained. The other part
contains 1,755 sequences from 188 subjects, which form
256 iris classes (note that not every individual provides iris
image sequences of both eyes). These sequences are
captured in three stages. In the first stage, three image
sequences of each eye are obtained. One month later, at
least two sequences of each eye are captured (often three or
four sequences per eye). In the third stage (i.e., three months
later), 30 out of 188 subjects provide 138 sequences again.
The total number of iris classes is thus 306 (2 25 þ 256).
Since all existing methods for iris recognition only use one
image for matching, we make use of the proposed scheme
for image quality assessment and selection described in
Section 3.1 to form an image set for algorithm comparison.
The resulting set includes 2,255 images corresponding to
306 different classes. Some samples from this set are shown
in Fig. 8.

4.2

Performance Evaluation of Image Quality
Assessment
In order to evaluate the performance of the proposed
algorithm for image quality assessment, we manually
collect 982 clear images, 475 motion blurred images,
431 occluded images, and 429 defocused images from the
CASIA Iris Database. One third of each image class are used
for training and the rest for testing. Fig. 9 shows the
distributions of the quality descriptor for different types of
images in training and testing stages and the two axes
respectively denote two feature components of the quality
descriptor (see (1) for definition).
From this figure, we can draw the following conclusions:
The clear images are well clustered and separated
from images from the other three classes, indicating
good discriminating power of the defined quality
descriptor. This is further confirmed by the results
shown in Table 2.
2. Severely occluded iris images are rich in middle and
high frequency components caused by the eyelashes,
which is an important factor in discriminating such
images from clear images. The results in Fig. 9
confirm this observation.
3. The quality descriptors of motion blurred images are
similarly distributed as those of defocused images.
The former include many high frequency components in the vertical direction inherently caused by
the scan mode of the CCD camera, whereas the latter
are governed by low frequency components. Since
they all lack middle frequency components, the
corresponding ratios of middle frequency power to
other frequency power are close.
Table 2 illustrates the classification results for the
training and testing samples. The results clearly demonstrate the effectiveness of the proposed scheme for image
quality assessment. Both Daugman’s method [18] (measuring the total high frequency power of the Fourier spectrum
of an iris image) and the method by Zhang and Salganicoff
[31] (detecting the sharpness of the pupil/iris boundary)
concentrate on assessing the focus of an iris image.
1.

TABLE 2
Classification Results

1528

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 12,

DECEMBER 2003

TABLE 3
Recognition Results Using Different Similarity Measures

However, our algorithm can discriminate clear images from
not only defocused images but also from motion blurred
images and severely occluded images. The results indicate
that the proposed scheme should be highly feasible in
practical applications.

4.3 Experimental Results of Iris Recognition
For each iris class, we choose three samples taken at the first
session for training and all samples captured at the second
and third sessions serve as test samples. This is also
consistent with the widely accepted standard for biometrics
algorithm testing [3], [4]. Therefore, there are 918 images for
training and 1,237 images for testing (To satisfy requirement of using images captured in different stages for
training and testing, respectively, 100 images taken at the
first session are not used in the experiments). Computing a
point estimation of a performance measure has been
extensively adopted in pattern recognition and is also easy
to use. Here, it is necessary to briefly introduce how we
obtain the interval estimation of a performance measure
using the bootstrap. In our experiments, the CRR, FMR, and
FNMR are three performance measures. A major assumption of the bootstrap for estimating an unknown distribution F is that the observations fx1 ; x2 . . . xn g in an original
sample x are independent and identically distributed. But
often, the case is just the contrary. With the FNMR as an
example, if more than one matching pair per iris class is
available in the known sample (i.e., at least two test samples
per class), the observed data is dependent. Moreover,
the bootstrap demands sampling with replacement from
n observations of the known sample to form thousands of
bootstrap samples. This implies that the observed data in a
bootstrap sample is a subset of the original observations. To
satisfy these two requirements of the bootstrap, we compute
confidence intervals of performance measures as follows:
1.
2.

3.

4.
5.

Construct a template set including 306 different
classes using 918 iris images.
Create a test set containing 306 iris classes by
drawing known iris classes with replacement. This
means that one iris class likely appears multiple
times in the test set.
For each iris class in the resulting test set, only one
test sample is chosen at random from all available
test samples of this class.
Compute the CRR, FMR, and FNMR using the
constructed template and test set.
Repeat Steps 2, 3, and 4 5,000 times and then,
respectively, estimate the 95 percent confidence
intervals of the CRR, FMR, and FNMR by the
percentile method [40].

Fig. 10. Recognition results using features of different dimensionality.

The percentile method establishes a ð1 2aÞ100 percent
confidence interval by computing the accumulated probability of a probability distribution from both sides. If the
accumulated probability exceeds a in l on left side and in u
on right side, the confidence interval is ½l; u . In the
following results, if the performance measure is denoted
by only a numerical value, this says that all 1,237 test
samples are used to estimate this measure. If the performance measure is expressed by a confidence interval, this
means that we adopt the bootstrap method described above
to calculate this interval.

4.3.1 Choice of Similarity Measures
Similarity measures play an important role in iris matching.
We thus perform a series of experiments to select a suitable
similarity measure for texture features generated by the
proposed method. Table 3 shows the recognition results
obtained with three typical measures based on the original
features and the dimensionality-reduced features, respectively. The dimensionality of the reduced feature vector is
200, whereas that of the original feature vector is 1,536.
As shown in Table 3, the three similarity measures lead
to very similar results when the original features are used
and the method’s performance does not vary drastically
after dimensionality reduction. This demonstrates that both
dimensionality reduction and similarity measures have
very small impact on recognition accuracy. The results also
show that the cosine similarity measure is slightly better
than the other two. Fig. 10 describes variations of the
recognition rate with changes of dimensionality of the
reduced feature vector using the cosine similarity measure.
From this figure, we can see that with increasing dimensionality of the reduced feature vector, the recognition rate
also increases rapidly. However, when the dimensionality
of the reduced feature vector is up to 150 or higher, the
recognition rate starts to level off at an encouraging rate of
about 99.27 percent. In particular, our method achieves a
recognition rate of 99.43 percent using only 200 features. In
the subsequent experiments, we utilize 200 features and the
cosine similarity measure for matching.
4.3.2 Recognition Results
We evaluate the proposed algorithm in two modes,
identification and verification. In identification tests, an
overall correct recognition rate of 99.43 percent is achieved
using all test samples and the corresponding 95 percent

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

TABLE 4
False Match and False Nonmatch Rates with
Different Threshold Values

confidence interval is [98.37 percent, 100 percent]. The black
lines of Fig. 13 show the verification results using the
ROC curve with confidence intervals. Table 4 lists three
typical system operating states in verification. In particular,
if one and only one false match occurs in 100,000 trails, one
can predict that the false nonmatch rate is less than
2.288 percent with a 95 percent confidence. These results
are highly encouraging and indicate high performance of
the proposed algorithm.

4.3.3 Performance Evaluation of the Defined
Spatial Filter
The spatial filter defined in Section 3.3.1 is thought to be
highly suitable for iris recognition since it is constructed
based on the observations about the characteristics of the
iris. This is confirmed by our experimental results shown in
Fig. 11. Filters used in our experiments include well-known
directional filters (i.e., Gabor filters in the horizontal
direction) and the defined spatial filters using the same
parameters as Gabor filters. Similar to the scheme for
showing the ROC curve with confidence intervals in [42],
we denote confidence intervals of the FMR and FNMR,
respectively. Fig. 11 shows the verification results of the
proposed method using different filters, which reveals that
the defined filters outperform Gabor filters. Gabor filters
only capture iris information in the horizontal direction,
whereas the defined filters not only show interest in
information in the horizontal direction but also consider
information from other directions. That is, the latter can

1529

obtain more information for recognition. The results also
show that one can achieve good results using Gabor filters
in the horizontal direction. This indicates that information
in the horizontal direction in a normalized iris is more
discriminating than that in other directions, namely, higher
information density in the angular direction in an original
image. We find from Fig. 11 that there is considerable
overlap in the FNMR confidence intervals on the two ROC
curves. This is because both the defined filters and Gabor
filters used in our experiments aim to capture discriminating iris information in the horizontal direction. The defined
filters, however, make effective use of more information in
other directions, which results in higher accuracy. The
above results demonstrate that iris information in the
angular direction has higher discriminability and information in other directions is a helpful supplement for more
accurate recognition.

4.3.4 Comparison with Existing Methods
The previous methods [17], [18], [19], [20], [21], [22], [23],
[24], [25], [26], [27], [28], [29], [30], [31] for iris recognition
mainly focus on feature representation and matching.
Therefore, we only analyze and compare the accuracy
and efficiency of feature representation and matching of
these methods. The methods proposed by Daugman [18],
Wildes et al. [20], Boles and Boashash [22] are probably the
best-known. They characterize local details of the iris based
on phase, texture analysis and zero-crossing representation
respectively. Here, we will present a detailed comparison
between the current method and their methods (and our
previous work) described in [18], [20], [22], [29], [30] on the
CASIA Iris Database. For the purpose of comparison, we
implement these methods according to the published
papers [12], [16], [17], [18], [19], [20], [21], [22]. Because
the method by Wildes et al. [20] only works in verification
mode, we do not test the performance of this method in
identification mode. Table 5 and Fig. 12 show the
identification results, and Fig. 13 describes the verification
results.

Fig. 11. Performance comparison of the defined filters and Gabor filters. The left and right plots show the 95 percent confidence interval (CI) of the
FMR and FNMR, respectively. The solid lines denote the bounds of the CI derived by the bootstrap and the dash-dot lines are the ROC curves based
on all test samples.

1530

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

VOL. 25, NO. 12,

DECEMBER 2003

TABLE 5
Identification Results

Using the bootstrap, we can approximately predict the
recognition rate distributions of these methods for a larger
test population. Fig. 12 shows the estimated distributions
for our previous methods, Boles’s method and the proposed
method. Table 5 gives the 95 percent confidence intervals of
these methods. Since Daugman’s method obtains 100 percent recognition rate, we do not plot its probability
distribution (the probability of 100 percent recognition rate
is 1) in Fig. 12 in order to express the differences of other
distributions more clearly. Looking at the results shown in
Table 5 and Fig. 12, we can find that Daugman’s method has
the best performance, followed by the proposed method
and the methods described in [30], [29], and [22], respectively. This conclusion is further consolidated by the
verification results given in Fig. 13.
Fig. 13 shows the ROC curves with confidence intervals
for these methods. The left and right plots show the
95 percent confidence interval (CI) of the FMR and FNMR,
respectively. The solid lines denote the bounds of the CI
derived by the bootstrap and the dash-dot lines are the

Fig. 13. ROC curves with confidence intervals.

Fig. 12. Distributions of correct recognition rates.

ROC curves using all test samples. Fig. 13 not only indicates
the performance of different methods but also provides
information of how much the performance of a given
method can vary. In general, the confidence interval of the
FMR is smaller than that of the FNMR since one can obtain
more nonmatching pairs (for estimating the FMR) than
matching pairs (for estimating the FNMR) in experiments.
We divide these methods into two groups and, respectively,
show the results in the top and bottom plots in Fig. 13 in
order to improve the legibility of the plots. The first group
includes [22] and [29], and the second [18], [20], [30] and the
proposed method. Two observations can be made from
Fig. 13. First, in terms of performance, a clear ranking for
these methods from best to worst is as follows: Daugman’s

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

method, the proposed method, our previous method [30],
Wildes’s method, our previous method [29], and Boles’s
method. Second, Boles’s method is close to our previous
method [29] and they are much worse than the other
methods. The above results show that the proposed method
is only inferior to Daugman’s method and much better than
the others.
Fig. 13 also shows that there is overlap in some of the
FNMR confidence intervals on the ROC curves, especially
among Daugman’s method, the proposed method, and our
earlier one [30]. As we know, the intraclass and the
interclass distance distributions jointly determine the
verification performance of an algorithm. That is, the
ROC curve is directly related to these two distance
distributions. In the experiments, we learned that large
matching scores from the intraclass distribution and small
nonmatching scores from the interclass distribution result
in the observed overlap in Fig. 13 (assume that the intraclass
distance is less than the interclass distance). Large matching
scores which correspond to false nonmatch pairs are mainly
caused by occlusions of eyelids and eyelashes, inaccurate
localization, and pupil movement (hence, iris distortion).
Small nonmatching scores which correspond to false match
pairs are decided to a greater extent by the inherent
discrimination ability of a method. If test samples include
some images which can lead to large matching scores or
small nonmatching scores, the ROC curve of a method will
inevitably deteriorate. Large matching scores can affect all
algorithms (our current comparison experiments do not
include eyelid and eyelash detection. To reduce false
nonmatching caused by large matching scores, we are
working on eyelid and eyelash detection, more accurate
localization and more effective normalization.). However,
small nonmatching scores depend on the specific iris
representation. The observed overlap indicates that the
discrimination ability of a given method may slightly vary
with the iris characteristics (hence, test samples). For
example, for iris images containing very few distinct
characteristics, the texture analysis based methods have a
relatively low accuracy. If test samples do not include such
images, both the proposed method and the method in [30]
can also achieve a quite high performance as shown in
Fig. 13. Based on the above analysis, we can conclude that
the range of a performance measure’s confidence intervals
reflects the corresponding method’s dependence on test
samples (the wider the range is, the stronger the dependence should be) and that the overlap in performance
measures’ confidence intervals among different methods
implies that the performance of these methods can be
almost the same on some specific test samples (the larger
the overlap is, the closer the performance). Considering
these two points, we can maintain the above performance
ranking for the six methods shown in Fig. 13.
Our previous method [29] divided the normalized iris
into eight blocks of the same size (64 64) and represented
the texture of each block using Gabor filters at five scales
and four directions. This implies that it captures less local
texture information of the iris. Boles et al. only employed
extremely little information along a concentric circle on the
iris to represent the whole iris, which resulted in a lower

1531

accuracy as shown in Fig. 13. Different from these two
methods, our current method captures much more local
information. Lim et al. [23] made use of the fourth-level
high frequency information of 2D Haar wavelet decomposition of an iris image for feature extraction. As we know,
the fourth-level details of an image’s wavelet decomposition contain essentially very low frequency information.
That is, their method did not effectively exploit middle
frequency components of the iris which play an important
role in recognition as well. Moreover, iris features in their
method are variant to eye rotation as the discrete wavelet
transform is not translation invariant. When there is a
rotation between a pair of irises from the same subject, their
method will generate a false nonmatch. Therefore, there is
no reason to expect that their method outperforms ours. The
proposed method is a significant extension of our earlier
algorithm presented in [30]. Compared with this earlier
method, the current method utilizes multichannel spatial
filters to extract texture features of the iris within a wider
frequency range. This indicates that the extracted features
are more discriminating, and the current method thus
achieves higher accuracy. Wildes et al. made use of
differences of binomial low-pass filters (isotropic Gaussian
filters) to achieve overall bandpass iris representation,
whereas we design more suitable spatial filters for recognition according to the iris characteristics. This leads to better
performance of our method. Furthermore, their method
relies on image registration and matching and is computationally demanding. In both identification and verification
tests, Daugman’s method is slightly better than the
proposed method. It should be noted that these two
methods explore different schemes to represent an iris.
We extract local characteristics of the iris from the viewpoint of texture analysis, whereas Daugman used phase
information to represent local shape of the iris details. In
essence, Daugman analyzed the iris texture by computing
and quantizing the similarity between the quadrature
wavelets and each local region, which requires that the
size of the local region must be small enough to achieve
high accuracy. Therefore, the dimensionality of the feature
vector (2,048 elements) in Daugman’s method is far higher
than ours (200 elements). That is, his method captures much
more information in much smaller local regions, which
makes his method better than ours.

4.4 Discussions and Future Work
Based on the above results and analysis, we can draw a
number of conclusions as well as find that some issues need
to be further investigated.
1.

The proposed scheme for iris image quality assessment is quite feasible. It can effectively deal with the
defocused, motion blurred, and occluded images.
The motion blurred images used in our experiments
are captured by an iris sensor working in the
interlaced scan mode. As stated in Section 3.1, the
motion blurred images taken by an iris sensor
working in the progressive scan mode are expected
to have similar frequency distribution to the defocused images. However, such an expectation need to
be further verified using real images.

1532

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE,

2.

3.

4.

5

The comparison experiment described in Section 4.3.3
reveals that information along the angular direction is
highly discriminating for recognition, which is consistent with the implicit reports by Daugman in [18].
We construct new multichannel spatial filters according to such a distinct distribution of the iris details and
the proposed method is thus expected to achieve a
satisfying recognition accuracy. In fact, our current
method achieves the best performance among the
existing methods based on texture analysis [20], [21],
[23], [28], [29], [30].
Table 5 and Figs. 12 and 13 show that the proposed
algorithm is only worse than Daugman’s method.
Increase of the dimensionality of the feature vector
improves the recognition accuracy of our method
but it does not outperform Daugman’s method. By
analyzing these two algorithms carefully, it is not
difficult to find that different viewpoints for feature
representation determine their differences in performance. Phase information characterized by Daugman reflects in essence local shape features of the
iris, whereas texture features used by us denote
statistical frequency information of a local region. By
careful examination on the appearance of numerous
iris images, we learn that a remarkable and
important characteristic of the iris is the randomly
distributed and irregular details. Local shape features are thus expected to better represent such iris
details than texture features. The experimental
results and analysis have indicated that local shape
features could be more discriminating iris features.
We are currently working on representing the iris
characteristics using the shape description method
in order to achieve higher accuracy.
The number and the class of iris samples used in our
experiments are of a reasonable size. Therefore, the
conclusions using the statistical bootstrap method
based on such a data set are useful for both research
and applications. We intend to expand our iris
database to include much more iris image sequences
from more races and make it publicly available to
promote research on iris recognition.

CONCLUSIONS

Biometrics based personal identification methods have
recently gained more interests with an increasing emphasis
on security. In this paper, we have described an efficientmethod for personal identification from an iris image
sequence. A quality descriptor based on the Fourier spectra
of two local regions in an iris image is defined to
discriminate clear iris images from low quality images
due to motion blur, defocus, and eyelash occlusion.
According to the distinct distribution of the iris characteristics, a bank of spatial filters is constructed for efficient
feature extraction. It has been proven that the defined
spatial filter is suitable for iris recognition. The experimental results have demonstrated the effectiveness of the
proposed method. A detailed performance comparison of
existing methods for iris recognition has been conducted on

VOL. 25, NO. 12,

DECEMBER 2003

the CASIA Iris Database. Such comparison and analysis will
be helpful to further improve the performance of the iris
recognition methods.

ACKNOWLEDGMENTS
The authors would like to thank Dr. John Daugman
(Cambridge University, UK), Dr. Richard Wildes (York
University, Canada), and Dr. Jianguo Zhang for their
helpful discussions. They also thank the anonymous
referees for their constructive comments. A public version
of the CASIA Iris Database is available from http://
www.sinobiometrics.com. This work has been filed for
patents and is funded by research grants from the NSFC
(Grant No. 69825105), the National 863 Program of China
(Grant No. 2001AA114180), and the CAS.

REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]

Biometrics: Personal Identification in a Networked Society, A. Jain,
R. Bolle and S. Pankanti, eds. Kluwer, 1999.
D. Zhang, Automated Biometrics: Technologies and Systems. Kluwer,
2000.
T. Mansfield, G. Kelly, D. Chandler, and J. Kane, “Biometric
Product Testing Final Report,” issue 1.0, Nat’l Physical Laboratory
of UK, 2001.
A. Mansfield and J. Wayman, “Best Practice Standards for Testing
and Reporting on Biometric Device Performance,” Nat’l Physical
Laboratory of UK, 2002.
F. Adler, Physiology of the Eye: Clinical Application, fourth ed.
London: The C.V. Mosby Company, 1965.
H. Davision, The Eye. London: Academic, 1962.
R. Johnson, “Can Iris Patterns Be Used to Identify People?”
Chemical and Laser Sciences Division LA-12331-PR, Los Alamos
Nat’l Laboratory, Calif., 1991.
A. Bertillon, “La Couleur de l’Iris,” Rev. of Science, vol. 36, no. 3,
pp. 65-73, 1885.
T. Camus, M. Salganicoff, A. Thomas, and K. Hanna, Method and
Apparatus for Removal of Bright or Dark Spots by the Fusion of
Multiple Images, United States Patent, no. 6088470, 1998.
J. McHugh, J. Lee, and C. Kuhla, Handheld Iris Imaging Apparatus
and Method, United States Patent, no. 6289113, 1998.
J. Rozmus and M. Salganicoff, Method and Apparatus for Illuminating and Imaging Eyes through Eyeglasses, United States Patent,
no. 6069967, 1997.
R. Wildes, J. Asmuth, S. Hsu, R. Kolczynski, J. Matey, and S.
Mcbride, Automated, Noninvasive Iris Recognition System and
Method, United States Patent, no. 5572596, 1996.
T. Tan, Y. Wang, and L. Ma, A New Sensor for Live Iris Imaging, PR
China Patent, no. ZL 01278644.6, 2001.
http://www.iris-recognition.org/, 2002.
L. Flom and A. Safir, Iris Recognition System, United Sates Patent,
no. 4641394, 1987.
J. Daugman, Biometric Personal Identification System Based on Iris
Analysis, United States Patent, no. 5291560, 1994.
J. Daugman, “High Confidence Visual Recognition of Persons by a
Test of Statistical Independence,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 15, no. 11, pp. 1148-1161, Nov. 1993.
J. Daugman, “Statistical Richness of Visual Phase Information:
Update on Recognizing Persons by Iris Patterns,” Int’l J. Computer
Vision, vol. 45, no. 1, pp. 25-38, 2001.
J. Daugman, “Demodulation by Complex-Valued Wavelets for
Stochastic Pattern Recognition,” Int’l J. Wavelets, Multiresolution
and Information Processing, vol. 1, no. 1, pp. 1-17, 2003.
R. Wildes, J. Asmuth, G. Green, S. Hsu, R. Kolczynski, J. Matey,
and S. McBride, “A Machine-Vision System for Iris Recognition,”
Machine Vision and Applications, vol. 9, pp. 1-8, 1996.
R. Wildes, “Iris Recognition: An Emerging Biometric Technology,” Proc. IEEE, vol. 85, pp. 1348-1363, 1997.
W. Boles and B. Boashash, “A Human Identification Technique
Using Images of the Iris and Wavelet Transform,” IEEE Trans.
Signal Processing, vol. 46, no. 4, pp. 1185-1188, 1998.

MA ET AL.: PERSONAL IDENTIFICATION BASED ON IRIS TEXTURE ANALYSIS

[23] S. Lim, K. Lee, O. Byeon, and T. Kim, “Efficient Iris Recognition
through Improvement of Feature Vector and Classifier,” ETRI J.,
vol. 23, no. 2, pp. 61-70, 2001.
[24] R. Sanchez-Reillo and C. Sanchez-Avila, “Iris Recognition With
Low Template Size,” Proc. Int’l Conf. Audio and Video-Based
Biometric Person Authentication, pp. 324-329, 2001.
[25] C. Sanchez-Avila and R. Sanchez-Reillo, “Iris-Based Biometric
Recognition Using Dyadic Wavelet Transform,” IEEE Aerospace
and Electronic Systems Magazine, pp. 3-6, Oct. 2002.
[26] C. Tisse, L. Martin, L. Torres, and M. Robert, “Person Identification Technique Using Human Iris Recognition,” Proc. Vision
Interface, pp. 294-299, 2002.
[27] T. Tangsukson and J. Havlicek, “AM-FM Image Segmentation,”
Proc. IEEE Int’l Conf. Image Processing, pp. 104-107, 2000.
[28] Y. Zhu, T. Tan, and Y. Wang, “Biometric Personal Identification
Based on Iris Patterns,” Proc. Int’l Conf. Pattern Recognition, vol. II,
pp. 805-808, 2000.
[29] L. Ma, Y. Wang, and T. Tan, “Iris Recognition Based on
Multichannel Gabor Filtering,” Proc. Fifth Asian Conf. Computer
Vision, vol. I, pp. 279-283, 2002.
[30] L. Ma, Y. Wang, and T. Tan, “Iris Recognition Using Circular
Symmetric Filters,” Proc. 16th Int’l Conf. Pattern Recognition, vol. II,
pp. 414-417, 2002.
[31] G. Zhang and M. Salganicoff, Method of Measuring the Focus of
Close-Up Images of Eyes, United States Patent, no. 5953440, 1999.
[32] K. Castleman, Digital Image Processing. Prentice-Hall, 1997.
[33] J. Daugman, “Uncertainty Relation for Resolution in Space, Spatial
Frequency, and Orientation Optimized by Two-Dimensional
Visual Cortical Filters,” J. Opticl Soc. of Am. A, vol. 2, pp. 11601169, 1985.
[34] T. Lee, “Image Representation Using 2D Gabor Wavelets,” IEEE
Trans. Pattern Analysis and Machine Intelligence, vol. 18, no. 10,
pp. 959-971, Oct. 1996.
[35] D. Clausi and M. Jernigan, “Designing Gabor Filters for Optimal
Texture Separability,” Pattern Recognition, vol. 33, pp. 1835-1849,
2000.
[36] A. Jain, S. Prabhakar, L. Hong, and S. Pankanti, “Filterbank-Based
Fingerprint Matching,” IEEE Trans. Image Processing, vol. 9, no. 5,
pp. 846-859, 2000.
[37] J. Zhang and T. Tan, “Brief Review of Invariant Texture Analysis
Methods,” Pattern Recognition, vol. 35, no. 3, pp. 735-747, 2002.
[38] P. Belhumeur, J. Hespanha, and D. Kriegman, “Eigenfaces vs.
Fisherfaces: Recognition Using Class Specific Linear Projection,”
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 7,
pp. 711-720, July 1997.
[39] C. Liu and H. Wechsler, “Gabor Feature Based Classification
Using the Enhanced Fisher Linear Discriminant Model for Face
Recognition,” IEEE Trans. Image Processing, vol. 11, no. 4, pp. 467476, 2002.
[40] J. Beveridge, K. She, B. Draper, and G. Givens, “A Nonparametric
Statistical Comparison of Principal Component and Linear
Discriminant Subspaces for Face Recognition,” Proc. IEEE Conf.
Computer Vision and Pattern Recognition, pp. 535-542, 2001.
[41] R. Bolle, S. Pankanti, and N. Ratha, “Evaluation Techniques for
Biometric-Based Authentication Systems (FRR),” IBM Computer
Science Research Report RC 21759, 2000.
[42] J. Wayman, “Confidence Interval and Test Size Estimation for
Biometric Data,” Nat’l Biometric Test Center (collectedworks), pp. 91101, 2000.
[43] B. Efron and R. Tibshirani, “Bootstrap Methods for Standard
Errors, Confidence Intervals, and Other Measures of Statistical
Accuracy,” Statistical Science, vol. 1, pp. 54-75, 1986.

1533

Li Ma received the BSc and MSc degrees in
automation engineering from Southeast University, China, in 1997 and 2000, respectively, and
the PhD degree in computer science from the
National Laboratory of Pattern Recognition,
Chinese Academy of Sciences, in 2003. Currently, he is a research member of IBM China
Research Lab. His current research interests
include image processing, pattern recognition,
biometrics, and multimedia.

Tieniu Tan (M’92-SM’97) received the BSc
degree in electronic engineering from Xi’an
Jiaotong University, China, in 1984, and the
MSc, DIC, and PhD degrees in electronic
engineering from the Imperial College of
Science, Technology and Medicine, London,
UK, in 1986, 1986, and 1989, respectively. He
joined the Computational Vision Group in the
Department of Computer Science at The University of Reading, England, in October 1989,
where he worked as a research fellow, senior research fellow, and
lecturer. In January 1998, he returned to China to join the National
Laboratory of Pattern Recognition, the Institute of Automation of the
Chinese Academy of Sciences, Beijing, China. He is currently a
professor and the director of the National Laboratory of Pattern
Recognition as well as president of the Institute of Automation. He has
published widely on image processing, computer vision, and pattern
recognition. His current research interests include speech and image
processing, machine and computer vision, pattern recognition, multimedia, and robotics. He serves as referee for many major national and
international journals and conferences. He is an associate editor of
Pattern Recognition and IEEE Transactions on Pattern Analysis and
Machine Intelligence, the Asia editor of Image and Vision Computing.
Dr. Tan was an elected member of the executive committee of the
British Machine Vision Association and Society for Pattern Recognition
(1996-1997) and is a founding cochair of the IEEE International
Workshop on Visual Surveillance. He is a senior member of the IEEE
and a member of the IEEE Computer Soceity.
Yunhong Wang received the BSc degree in
electronic engineering from Northwestern Polytechnical University, the MS degree (1995) and
the PhD degree (1998) in electronic engineering
from Nanjing University of Science and Technology. She joined the National Lab of Pattern
Recognition, Institute of Automation, Chinese
Academy of Sciences in 1998, where she has
been an associate professor since 2000. Her
research interests include biometrics, pattern
recognition, and image processing. She is a member of the IEEE and
the IEEE Computer Society.
Dexin Zhang received the BSc degree in automation engineering from Tsinghua University in
2000. Then, he joined the National Laboratory of
Pattern Recognition, Chinese Academy of
Sciences, to pursue his master’s degree. His
research interests include biometrics, image
processing, and pattern recognition.

. For more information on this or any other computing topic,
please visit our Digital Library at http://computer.org/publications/dlib.



Documents similaires


03 fourier ma
foreground background segmentation using temporal and spatial
lidar topographie
ms 33 424
2015021112wavelets
ijcai07 259